DigestAI news desk

Cut through the AI noise.

Agents & Tools6 min read

H company launches Holo4 agentic models for desktop and API tasks

H company announced the Holo4 series, offering two agentic models—Holo4‑27B (dense) and Holo4‑35B‑A3B (Mixture of Experts)—through its H Models API. An updated Holotron4 Nano is also released. The models can interact with software via GUIs, code, MCP and APIs, aiming at real business workflows rather than single‑interface tasks.

1 source primary source

Key points

  • Holo4 released in two sizes, 27B dense and 35B‑A3B Mixture of Experts, on H Models API.
  • On OSWorld 2.0 benchmark Holo4‑27B scores 61.7% versus Opus 5.5’s 81.8%, while Holo4‑35B‑A3B scores 30.9%.
  • Holo4 can click, type, write code and call APIs across desktop, web, Android and sandbox environments using a single model.

On the OSWorld 2.0 benchmark Holo4‑27B achieved a 61.7% score, trailing the closed‑source Opus 5.5 at 81.8%, while the larger Holo4‑35B‑A3B scored 30.9%. Despite lower scores, Holo4 delivers comparable performance at a much lower cost per task. The company also applied its stack to NVIDIA’s Nemotron 3 Nano Omni, producing the Holotron4 Nano model that improves on the base model across GUI, API and coding environments.

All model weights are available on Hugging Face in multiple formats (BF16, FP8, NVFP4, 4‑bit GGUF), and the company plans to release optimized DSpark drafter checkpoints soon, further accelerating inference for developers.

Model page: Holo4 →

Full story from Hugging Face primary sourceOpen source ↗

Holo4: powering generalist computer-use agents

Hugging Face · 28 September 2026

Loading the full article…

This text was published by Hugging Face. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗

Topics · follow one to build your own front page
H companyNVIDIAHolo4-27BHolo4-35B-A3BHolotron4 NanoNemotron 3 Nano Omni

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.

Comments

via GitHub Discussions

More in Agents & Tools

All →

Related stories